A depth completion method based on cross-feature extraction window

A Dense depth maps with clear object boundaries are very important for robot mapping and navigation. However, today’s depth sensors can only produce sparse depth data, so it is necessary to complete them. Traditional methods mostly rely on RGB images as an auxiliary modality to provide color information in the depth-filling process. However, RGB by itself is not enough to provide all the required color information, so we add semantic maps as a guide on this basis. In addition, we add a cross-feature extraction window (CFEW) based on the attention mechanism to the feature extraction stage of this model. The essence of the window is to use the channel attention mechanism, the spatial attention mechanism, and the mutual features attention mechanism to enhance the important features and suppress the irrelevant features, which can enhance the border of the object in the environment.